Most guides on this topic end at the same place: upload a scan, press one button, download something shinier. That works for a faded holiday snapshot. It falls apart the moment the photo has a torn corner across someone's cheek, silver mirroring across half the emulsion, or a face that only survives as forty pixels of grey. At that point the model is not repairing anything, it is guessing, and the quality of the guess is the whole job. The tool roundups we keep running into, including our own comparison of AI photo enhancement tools, rarely say that part out loud.
This piece is about the parts that decide whether a restoration looks like your grandmother or like a stranger wearing her coat. Damage repair, face reconstruction, denoising, colorization and upscaling are different problems, and models good at one are usually mediocre at another. Running them as separate passes, in the right order, beats any single one-click restore.
It also assumes you care about accuracy. A restored family photo is a record, not a render. If you are working in a general AI photo editor and the output looks plausible but nobody in the room recognises the person, the restoration has failed even though the image is technically cleaner.
The Scan Decides More Than the Model
The single biggest quality jump in old photo restoration is not a model choice, it is the input. A phone snapshot of a print taken at an angle under a ceiling light carries glare, keystone distortion, and a colour cast baked into every pixel. No restoration model removes those cleanly, because it cannot tell the difference between a lighting artefact and the photograph itself.
Scan flat, at 600 dpi minimum for a standard print, 1200 dpi if the original is small or you intend to enlarge. Turn off the scanner's own dust removal and auto colour correction, both of which smear grain and shift skin tones before the AI ever sees the file. Save as TIFF or PNG, never JPEG, since compression artefacts get amplified by every later pass, exactly the way they do when you upscale a low quality image through an API.
A few practical notes from doing this on a shoebox of 1950s prints:
- Glass-plate negatives and slides should be scanned with transmitted light, not reflected, or you lose most of the tonal range
- Photographs stuck to album pages should be scanned in place; peeling them off destroys more than any model can rebuild
- Curled prints flatten under glass, but keep the glass clean or you add a second layer of dust to remove
If your only copy is already digital and low resolution, start with a dedicated AI image upscaler rather than asking a restoration model to enlarge and repair at the same time. Combined passes tend to invent texture where they should be recovering it.
What AI Actually Repairs, and What It Fabricates
It helps to sort damage into two categories. The first is damage where the surrounding pixels contain enough information to reconstruct what was lost: scratches, dust, small tears, fading, mild blur, uneven exposure. Inpainting models handle these well because they are interpolating, not imagining.
The second is damage where the information is genuinely gone: a missing corner containing an entire person, a face burned out by light damage, a body torn away. Here the model produces something statistically likely rather than something true. The same distinction applies to any subtractive edit, which is why the better object removal tools are judged on whether the fill matches surrounding context rather than on how invisible the seam is.
Run It as a Pipeline, Not a Single Pass
The sequence that consistently works: repair physical damage first, then denoise, then handle faces, then colorize if you want colour, then upscale last. Upscaling early locks in errors at a larger size, and colorizing before damage repair spreads colour into scratches. Chaining these steps by hand across four browser tabs gets tedious quickly, which is why people building repeatable versions of this tend to wire the steps together in an AI workflow tool where each model's output feeds directly into the next.
The order matters more than the specific models. Whatever you use, keep every intermediate file, because you will want to go back two steps when the face pass overcooks someone's eyes. Most editor and photo tool roundups rate one-click output and never test whether a tool lets you keep the steps in between.
- Damage pass - inpaint scratches, tears, stains, and emulsion loss with a mask, not a global filter
- Denoise pass - light touch only; aggressive denoising turns film grain into plastic
- Face pass - run face restoration separately at a low strength, and compare against the original crop
- Colour pass - optional, and the step most likely to introduce fiction
- Upscale pass - last, once the pixels you are enlarging are actually correct
If you have more than a handful of photos, this becomes a batch problem rather than an editing problem, and the same logic behind batch image generation via API applies: fix the pipeline on three representative images, then run the rest through it unattended and review the outputs as a set.
Faces Are Where Restorations Go Wrong
Face restoration models were trained overwhelmingly on modern, well-lit, high-resolution portraits. Point one at a 1930s wedding photo and it will pull the face toward that training distribution: smoother skin, more symmetric features, modern-looking eyes, teeth that were never visible in the original. The effect is subtle enough to miss on screen and obvious to anyone who knew the person. This is the same failure mode that makes AI headshot tools produce photos that look like a better-looking cousin of the subject.
Two defences work. Run the face pass at partial strength and blend it back over the original rather than replacing it outright, and always view the result at 100 percent next to the untouched crop. If the jawline moved, the restoration went too far.
For very small faces, accept a softer result. A blurry but honest face reads as an old photograph; a sharp invented one reads as a modern render, the same uncanny quality you see in realistic AI photo generators when the lighting is right but the identity is nobody's.
Colour Is a Guess, So Label It
Colorization models do not know what colour the dress was. They know what colour dresses usually are in their training data. The output is a plausible reconstruction, not evidence.
Two habits keep this honest: keep and share the monochrome version alongside the colorized one, and correct the colour manually where you have outside knowledge, such as a uniform, a car, or a house that still exists. Most AI image editors let you paint colour hints into specific regions before the model runs, which produces far better results than accepting the default pass.
Keep the Original, and Keep the Recipe
Archive the raw scan untouched, then keep the restored version as a derivative with a short note on what was repaired and what was reconstructed. Anyone who inherits the files later will thank you. If you rebuild the same pipeline often, saving it as a reusable graph in a node-based AI canvas means the recipe survives alongside the images instead of living in your browser history.
Restored stills also open up things the original could not do. Once a photograph is clean and correctly upscaled, it becomes a usable source for further work, including animating still images for a family archive or a memorial video.
FAQ
Can AI restore a photo that is torn in half?
It can rebuild the seam convincingly if both halves survive, because the model has context on either side. If one half is missing entirely, whatever fills the gap is invention, and should be treated as such.
What resolution should I scan old prints at?
600 dpi for prints you plan to view and share, 1200 dpi for small originals or anything you intend to enlarge. Scanning higher than the print's own detail level adds file size without adding information, though it does give upscaling models a cleaner starting point.
Do I need to colorize a black and white photo?
No, and often you should not. Monochrome is the original medium, and colorization is the least accurate step in the pipeline. Do it as a second version if you want it, and keep the original.
Why do restored faces look like different people?
Face restoration models pull faces toward modern training data, so identity drifts when the input is small or heavily damaged. Lower the strength, blend with the original, and compare crops side by side, the same discipline that keeps AI image editing tools from over-processing a portrait.
Can I restore hundreds of photos at once?
Yes, but fix your settings on a small representative sample first. Batch runs amplify whatever the pipeline gets wrong, and faces in particular need spot checks rather than blanket approval.
Is free AI photo restoration good enough?
For light fading and dust, usually yes. For heavy damage, free tools tend to apply one aggressive global pass and produce a smoothed, plastic result, which is the main difference the paid and self-hosted options are selling, as the enhancement tool breakdown covers in more detail.
What file format should I keep the restored version in?
TIFF or PNG for the archive copy, JPEG only for sharing. Every re-save of a lossy file degrades the work you just did.
Wrapping Up
AI has made photo restoration accessible to anyone with a scanner, but the one-click framing hides the part that matters. Split the job into passes, keep the scan clean, run faces conservatively, treat colour as interpretation, and archive the original next to the result. What you get is a photograph that still belongs to the person in it, and a repeatable process you can point at the rest of the box, or reuse later when you want to make an old photo talk or move.